scRNA-seq Dataset Integration

​

Vinicius Maracaja-Coutinho

​

​

Associate Professor

Universidad de Chile

vinicius.maracaja@uchile.cl

Modified version of Erick Armingol slides

Increasing complexity in single-cell RNA-seq data

(PMID: 37198436)

Building reference atlases

Hrovatin et al., Unpublished.

Usage of integrated databases

Hrovatin et al., Unpublished.

Using the available transcriptomes

(Hemberg Lab: https://www.singlecellcourse.org/)

Label-centric

comparison

Comparing the labels or annotations of cell types or clusters across different datasets or conditions.

​

• How well the cell types or clusters identified in one dataset match or align with those in another dataset after integration.

​

• Can be used to compare the annotations of two different samples from the same experiment.

​

• Can project cells from a new experiment onto an annotated reference.

Using the available transcriptomes

(Hemberg Lab: https://www.singlecellcourse.org/)

Cross-dataset

analysis

Attempts to computationally remove experiment-specific technical/biological effects so that data from multiple experiments can be combined and jointly analyzed.

Integration and analysis of multiple datasets

(PMID: 37198436)

Main challenges: Batch effects among datasets

Single-cell colored by datasets 🡪 Separation by dataset

Batch effect could be defined as variability in the data that is not due to a variable of interest

• Technical variability:
• Sample handling
• Experimental protocols
• Sequencing platform

​

• Biological variability:
• Donor differences
• External factors (e.g. environmental factors)
• Evolution

(PMID: 31217225)

Main challenges: Batch effects among datasets

Single-cell colored by datasets 🡪 Separation by dataset

Batch effect could be defined as variability in the data that is not due to a variable of interest

• Technical variability:
• Sample handling
• Experimental protocols
• Sequencing platform

​

• Biological variability:
• Donor differences
• External factors (e.g. environmental factors)
• Evolution

(PMID: 36859475)

Identifying covariates driving batch effects help to correct

(PMID: 31217225)

Main types of methods for data integration

(PMID: 36859475)

1. Linear decomposition methods

​

• Similarity-based batch correction methods
• Dimensionality reduction
• Identification of similar cells between batches

- Cell-level similarity search

- Cluster-level similarity search

​

3. Generative models with variational autoencoder (artificial intelligence)

​

Main types of methods for data integration

(PMID: 36859475)

Linear Decomposition

(PMID: 37002403)

Similarity: Mutual Nearest Neighbors (MNN)

(Hemberg Lab: https://www.singlecellcourse.org/)

Similarity: Mutual Nearest Neighbors (MNN)

(Hemberg Lab: https://www.singlecellcourse.org/)

Similarity: Mutual Nearest Neighbors (MNN)

(Hemberg Lab: https://www.singlecellcourse.org/)

Seurat-v3 (Canonical Correlation Analysis + Anchors)

(PMID: 31178118)

Harmony

(PMID: 31740819)

Deep-Learning models

(PMID: 36859475)

Benchmarking